The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python and Bash scripts to execute Git commands (e.g., git status, git rev-parse, git log) for repository auditing and routine synchronization. These commands are executed using argument lists, which prevents shell injection vulnerabilities. The instructions also emphasize using the least mechanism necessary, such as natural language instructions over shell hooks.
[PERSISTENCE]: The init_session_start_hook.py script manages Claude Code startup hooks in .claude/settings.json. The implementation includes defensive measures such as atomic writing, idempotent updates, and validation to prevent shell-sensitive character injection in guide paths. It also provides a clear removal mechanism for these managed entries.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes external repository data to infer setup requirements.
Ingestion points: The skill reads manifest files (package.json, uv.lock), project guides (ONBOARDING.md), and Git history logs.
Boundary markers: While explicit text delimiters are not mandated for all inputs, the instructions emphasize treating inferred data as evidence to be verified by smoke tests.
Capability inventory: The skill can execute Git commands and write to local configuration files.
Sanitization: Input validation in the hook manager prevents path traversal and shell injection when processing repository-relative paths.
[DATA_EXPOSURE_AND_EXFILTRATION]: The sanitize_history.sh script is a defensive utility that scans the repository's Git history for potentially hardcoded secrets (e.g., API keys, PATs). It is read-only by default and provides remediation guidance rather than performing automatic history rewrites.

auto-repo-setup